Cloud-computing giant Alibaba has developed a new chip that is more versatile than its older chips, the Wall Street Journal reported.
This comes as the Chinese cloud and AI firms are accelerating a push to develop domestic semiconductors and related technology after US restrictions and diplomatic friction disrupted access to advanced foreign chips.
Alibaba’s new chip, now in testing and described by people familiar with the project, departs from its earlier task-specific processors by targeting a broader set of inference use cases. The move reflects a broader industry pivot: cloud providers and start-ups in China are engineering substitutes for foreign processors such as Nvidia’s H20 after regulatory and trade tensions have complicated that supply channel.
Start-Ups & Incumbents Racing to Adapt
A number of Chinese firms have announced or fast-tracked chips this year. Shanghai-based MetaX unveiled a two-chip solution it says can substitute for the H20 on some workloads by combining smaller dies to boost effective memory capacity, even if it consumes more power.
Beijing’s Cambricon reported a strong quarter on brisk Siyuan 590 orders, and Huawei continues to push its Ascend line, at times claiming comparable performance on certain metrics by aggregating many chips into large systems.
Beijing has backed the industry with large funds and policy support, including an $8.4 billion AI-investment fund announced earlier this year, to accelerate self-sufficiency.
Yet Chinese manufacturers still face material constraints: fabs here largely rely on older tooling and non-leading-edge processes because US restrictions limit access to the most advanced chipmaking equipment. That complicates efforts to scale capacity quickly and to produce chips capable of high-end model training.
Inference vs Training: Where is China’s Focus
As per the WSJ report, Alibaba’s design stresses that the new processor is optimized for inference, running trained models to generate outputs, rather than the far more demanding task of training state-of-the-art foundations.
Training typically requires the very top-tier GPUs and interconnects that remain difficult to procure at scale in China, keeping a gap between domestic offerings and the world’s fastest training infrastructure.
One notable design choice: Alibaba’s chip aims to be compatible with software written for Nvidia platforms, potentially lowering switching costs for engineers who have built tools and workflows around Nvidia stacks. That contrasts with some Huawei chips, which are less compatible and therefore have seen limited uptake among private cloud operators wary of deepening dependence on a direct competitor.
Geopolitics & Market Signals
Regulatory whiplash has amplified urgency. In July, the US eased some export constraints so Nvidia could resume H20 shipments to China, but Chinese authorities subsequently discouraged purchases, keeping demand muted.
Meanwhile, start-ups such as DeepSeek have signalled that software innovations combined with improved domestic silicon could narrow the performance gap sooner than many expect, a claim that has helped fuel local investor enthusiasm.
Analysts say combining many chips or leaning on older process nodes can produce practical performance improvements for inference and specific enterprise use cases, and that China’s push will yield valuable alternatives and more diversified supply chains.
Still, closing the divide with leading-edge US chips for large-scale model training will remain a longer-term challenge because of tooling, heat and reliability constraints. For now, Alibaba and others are seeking to carve out commercially useful, China-made compute for the booming inference market while the industry and policy landscape continue to evolve.
